The dynamic layout problem: improved genetic algorithms

Jaydeep Balakrishnan1
Chun Hung Cheng2

Abstract

In this paper, we examine the use of genetic algorithms for the dynamic layout problem. Two
versions of genetic algorithms are proposed. Our genetic algorithms differ from an existing
implementation in two ways: first, we adopt a simple crossover operator to avoid illegal offspring,
and second, we use a new generational replacement strategy to help increase population diversity.
Computational study shows that one version of our implementations outperforms an existing
approach for the problem.

1. Faculty of Management, University of Calgary, Calgary, Alberta T2N1N4, Canada.

Ph: (403) 220 7844 Email: [email protected]
2. Department of Systems Engineering and Engineering Management, The Chinese University of
Hong Kong, Shatin, Hong Kong. Ph: +852 2609 8322 Email: [email protected]
 The dynamic layout problem: improved genetic algorithms

Abstract

In this paper, we examine the use of genetic algorithms for the dynamic layout problem. Two
versions of genetic algorithms are proposed. Our genetic algorithms differ from an existing
implementation in two ways: first, we adopt a simple crossover operator to avoid illegal offspring,
and second, we use a new generational replacement strategy to help increase population diversity.
Computational study shows that one version of our implementations outperforms an existing
approach for the problem.

1
 The dynamic layout problem: improved genetic algorithms

1. Introduction

This paper investigates the design of facility layouts based on a multi-period planning horizon.

During this horizon, the material handling flows between the different departments in the layout may

change. This necessitates a more sophisticated approach than the much researched static plant layout

problem (SPLP) approach. The dynamic plant layout problem (DPLP) extends the SPLP by assuming

that the material handling flows can change over time. This in turn might necessitate layout

rearrangement.

In the static problem, given a group of departments, material flow between each pair of

departments, and the cost per unit of flow per unit distance, we have to arrange the departments into a

layout such that the sum of the costs of flow between departments in the layout is minimised. The

material flows between pairs of departments or relative material flows are assumed to be constant. In an

environment where flows do not change over a long period of time, this assumption is justified. But in

today's market based and dynamic environment, such flows can change quickly. Page (1991) reports

that on average, 40% of a company’s sales come from new products. In the dynamic environment, we

can still use the static approach in two ways. The first is to use a short planning horizon so that during

this horizon the flows are fairly constant. The disadvantage is that after the short horizon, if our relative

flows change, we have to rearrange the layout or if we do not, be constrained by an inefficient layout.

Rearranging layouts frequently without prior planning can result in operational disruptions and excess

rearrangement costs. The second approach is to use a long planning horizon and disregard the changes

in flow. The total flow over the planning horizon can be used assuming constant flow. There will be no

rearrangement costs, but this may result in the layout being inefficient throughout the horizon.

The dynamic approach to layout corrects the above deficiency. In the dynamic approach, we

plan our layout based on a multi-period time horizon. During this time if the material flow changes

warrant it, we may plan for one or more layout rearrangements. In this research, an extension is made to

2
the dynamic approach. It is assumed that there is a constraint on the funds available for layout

rearrangement. In addition to its effect on layout planning, the constrained dynamic layout approach has

broader implications on the marketing and finance functions of the firm.

Any change in product mix can result in changes in flow and thus affect layouts. Layouts

designed on the basis of short term demand or flows can result in excessive material handling costs if

the product mix changes. The importance of good layout planning can be gauged from the fact that over

$250 billion is spent in the U.S. alone on layouts that require planning and replanning and that 20% to

50% of the total operating expenses within manufacturing can be attributed to material handling

(Tompkins & White, 1984). The next section explains the dynamic problem through an example

adopted from Rosenblatt (1986).

A rectangular layout with six locations is shown in Figure 1. The locations are all equal in size.

There are six departments that have to be assigned to the six locations. One such assignment is shown in

the figure and is designated by the sequence 1-2-3-4-5-6.

1 2 3

4 5 6

Figure 1: An example problem

This is only one out of the 6! or 720 combinations that exist for this layout. Each combination

represents a different static layout. In this particular case, due to symmetry only 120 different static

solutions or combinations exist. But in the dynamic problem, all the 720 different solutions would have

3
to be examined for simplicity. We are interested in determining the assignment of departments in this

layout over a five period planning horizon. A period could represent a year, quarter, month, etc.

The varying relative flows between pairs of departments along with shifting cost for each

department is shown in Figure 2. The shifting cost depends only on the department shifted and not on

the distance of the move. This assumption is valid where the fixed costs of the move such as removing

the machine and re-installing it, lost cost of lost production dominate the variable costs such as the cost

to move the machine a unit distance. The layout also shows flow dominance. Some of the departments

have higher material handling inflows than the others. These flow dominant departments change from

period to period. This ensures that the optimal static layout also changes during the planning horizon.

Under these varying conditions, it may be worthwhile to change over to a new layout one or more times

during the horizon. If it did not cost us anything to shift departments, the optimal solution would have

been to shift to the optimal static solution in each period. But since shifting departments costs money,

we have to balance the shifting costs with the savings in the cost of material handling due to this

shifting.

So the dynamic problem involves selecting a static layout for each period and then deciding

whether to change to a different layout in the next period. If the shifting costs are relatively low, we

would tend to shift or change the layout configuration more often in order to save flow costs. The

reverse is true for high shifting costs. The optimum solution to the dynamic problem might not involve

any of the static optimum solutions. Figure 3 gives us an appreciation of how difficult it really is to find

this optimum.

4
 Period 1

To -- 1 2 3 4 5 6
From
1 0 396 229 280 294 2498
2 2590 0 243 304 329 2237
3 2244 468 0 299 336 1875
4 2373 369 207 0 329 3021
5 2456 362 176 198 0 2130
6 2484 455 267 204 326 0

Period 2
1 0 2893 2227 341 353 449
2 379 0 2306 283 387 319
3 283 2371 0 338 318 351
4 295 2066 2361 0 303 486
5 385 2509 2370 278 0 425
6 421 2098 1708 349 331 0

Period 3
1 0 2161 201 1428 448 300
2 478 0 290 1752 400 444
3 426 3171 0 1774 421 323
4 431 3103 261 0 528 332
5 352 2821 219 1794 0 361
6 338 2717 218 2035 456 0

Period 4
1 0 422 377 1707 1288 1866
2 443 0 318 1831 1260 1265
3 325 337 0 1888 1565 1875
4 305 470 301 0 1644 1848
5 324 343 275 1852 0 2064
6 417 334 233 1284 1372 0

Period 5
1 0 2777 419 479 368 350
2 1372 0 557 427 411 370
3 1958 2886 0 460 404 469
4 1803 2558 372 0 265 470
5 1875 2526 353 414 0 392
6 1454 2570 432 419 373 0

Shifting cost for departments

1595 820 1725 1369 697 1188

Figure 2: Material Flow and Shifting Costs

Period

5
 1 2 3 4 5

Layout 1 123456 123456 123456 123456 123456

2 123465 123465 123465 123465 123465

3 123645 123645 123645 123645 123645

Layout 4 ... ... ... ... ...

Layout 5 ... ... ... ... ...

...

719 654312 654312 654312 654312 654312

720 654321 654321 654321 654321 654321

Figure 3: The Dynamic Layout Problem

Layouts are denoted by numbers, e.g. 123456 represents the layout shown in Figure 1. Let us

assume that our initial layout (in the present period, Period 1) is 123456 (Layout 1). Our initial layout

could have been any of the (n!) or 720 possible layouts in the period where n is the number of

departments in the layout which in our case is six. In the second period we have the option of changing

to any of the other 719 layouts or keeping the same layout. This analysis can be repeated for each of the

720 different layouts in each of the five periods. In other words, we would have to explicitly or

implicitly evaluate (n!)t combinations where 't' is the number of periods, which even for a six

department, five period problem is 1.93*1014 combinations, a very large number indeed! So the

6
dynamic facility layout problem can become extremely difficult to solve to optimality for even small

problems. Thus most practical sized problems will have to be solved heuristically, i.e. not all the

possible layouts in a period are included in the dynamic problem. But now, we have no guarantee of

optimality for the whole problem. Rosenblatt uses dynamic programming to solve the problem with

each layout in each period being a state and each period a stage.

Despite the importance of demand changes, little research has been done on the layout

problem under the condition of changing demand (i.e., the dynamic layout problem). A few heuristic

approaches are available to solve this complicated combinatorial problem. In the next section, we

will review these approaches. A recent approach by Conway and Venkataramanan (1994) makes use

of genetic algorithms (GAs). In this paper, we propose two versions of GAs for efficiently solving the

problem. Our GAs differ from the Conway and Venkataramanan (1994) GA in two aspects:

· We employ a simple and unique crossover operator to avoid illegal offspring.

· We use a different generational replacement approach to increase population diversity.

Our computational results show that one version of our GAs outperforms better than the Conway and

Venkataramanan (1994) GA in terms of the quality of solutions.

2. Literature Review

Solving the dynamic layout problem (also called the dynamic plant layout problem) includes

solving the static layout problem. Therefore this section includes a short discussion of research in

static layout and then a detailed discussion of dynamic layout.

The static layout problem can be formulated as a Quadratic Assignment Problem (QAP),

(Koopmans and Beckman, 1957). It is termed so because the objective function is a product of two

zero-one variables. Sahni & Gonzalez (1976) show this to be a NP-complete problem. Solving

problems with more than fifteen departments to optimality is very difficult. Thus it is clear that for most

practical size layouts, where a large number of departments are present, obtaining the optimal solution

7
will not be possible. So various heuristic algorithms have been proposed which can solve fairly large

size problems in reasonable amounts of time.

Heuristic algorithms can be classified into two major types - construction type algorithms where

a solution is constructed from scratch and, improvement type algorithms, where an initial layout is

improved. CRAFT (Armour & Buffa, 1963) , one of the more popular algorithms is an improvement

algorithm and is used in this research. For a comprehensive review of the static layout literature , see

Kusiak and Heragu (1987).

Less literature is available in the dynamic facility layout area. Solutions approaches to dynamic

facility layout are based on the dynamic location literature. Ballou (1968) considers the dynamic nature

of demand in locating warehouses. Location decisions are generally made on the basis of horizons of

twenty years or more. Thus the changes in demand during that time (dynamic nature) can be important.

If accurate predictions of the longer term demand during that horizon are available, then it may be

worthwhile to consider a more sophisticated analysis than simple static facility location. Dynamic

location analyses allow us to foresee layout changes thus ensuring that our plans will not be outdated by

the time it is implemented. Ballou presents an example of a manufacturing plant and a set of demand

points that shift over time periods. The optimal warehouse location for each period is different if we

considered that period's demand in isolation. But if we decide to shift the warehouse every period, we

would incur shifting costs. But on the other hand, the distribution costs would potentially be lower than

had we not shifted. So the objective here is to trade-off the costs of relocation versus the distribution

costs saved. Ballou uses Dynamic Programming (DP), proposed in Bellman and Dreyfus (1962), to

solve the multi-period model. Each stage corresponds to a period and at each stage the states are

different static location decisions. Ballou uses only the optimal solutions of each period as the states for

a period. Thus, in his five period example, each stage consists of five states: the five static solutions.

Since not all the static solutions are included, the DP solution is actually heuristic (sub-optimal) with

respect to the whole problem.

8
 Lodish (1970) shows that the DP solution time could explode for practical sized problems since

the optimal solution to the dynamic facility location problem would have to include all possible static

solutions. Sweeney and Tatham (1976) provide a fathoming procedure for the dynamic facility location

problem, that allows us to eliminate many of the states required in each period (stage). This makes the

DP computation easier.

Rosenblatt (1986) applies the Sweeney and Tatham model to the dynamic facility layout

problem. The problem is similar to dynamic warehouse location. The trade-off here is between the flow

(material handling) cost within the facility versus the shifting costs for the departments we might want

to relocate within the facility. One difference between the Sweeney and Tatham model and Rosenblatt

is that the shifting costs in the latter problem depend on the department being relocated whereas in the

former, it is always constant. The number of possible layouts in each period is given by n! where 'n' is

the number of departments. If a horizon of 't' periods are considered , then the total number of possible

combinations to be evaluated is (n!)t. So the problem can explode very easily in the dynamic facility

layout also. Rosenblatt also uses DP to solve a six department problem. He reports that in the tests

conducted in his experiment, the fathoming procedure of Sweeney and Tatham did not help in

eliminating any state.

Balakrishnan et al. (1992) extends the dynamic layout problem of Rosenblatt. They consider the

existence of budget constraints and conduct a detailed experiment to investigate the problem. Layout

rearrangement requires funds and these funds may be limited. So the dynamic layout problem is solved

under this constraint. The constrained shortest path (CSP) algorithm of Mote et al. (1988) is used to

solve the problem. In addition an experimental design involving the number of static layouts, the

method of static layout selection, constraint tightness and number of departments was undertaken. The

results show that CRAFT is a good method to generate static layouts and that using some random static

layouts is useful. The CSP was also shown to be more effective than two state variable dynamic

programming in solving this problem.

9
 Balakrishnan (1993) also proposes a fathoming procedure to reduce the number of possibilities

in dynamic facility layout. It states that any static layout, that differs in cost from the optimal static

layout by more than twice the maximum shifting cost, can be eliminated. One advantage of this method

is that this can be applied before solving any of the static layouts whereas other fathoming methods

require static layout solutions before application. The effectiveness of this fathoming method depends

on the magnitude of the rearrangement costs.

Urban (1993) proposes a steepest-descent pairwise-interchange heuristic for the dynamic layout

problem. This procedure combines the static and dynamic layout problem into one. Many other

methods separate the two. Varying planning horizons are considered. The results indicate that it

performs as well as existing methods and only slightly worse than optimal. The computational

requirements are also quite low for this method whereas for dynamic programming it could be

substantial.

Conway and Venkataramanan (1994) examines the suitability of genetic algorithms to the

problem. A genetic search employ the mechanics of natural selection and natural genetics to evolve a

population of initial solutions into a near-optimal solution. This approach is suited to handle multiple a

non-linear objective function as well as side constraints. In tests, it performed well compared to

dynamic programming for 6 and 9 department problems.

10
3. The dynamic layout problem

The dynamic layout problem is formulated as an extension of the static layout problem.
P N N N P N N N

Min 
t 2 i 1 j 1 l 1
AtijlYtijl + C
t 1 i 1 j 1 l 1
XtijXtkl
tijkl (1)

s.t. X
j 1
tij = 1, i = 1, …, N , t = 1, …, P (2)

X
i 1
tij = 1, j = 1, …, N , t = 1, …, P (3)

Ytijl = X(t-1)ij ´ Xtil, i, j, l = 1, …, N, t = 2, …, P (4)

where
i, k : Departments in the layout
j,l : Locations in the layout
Ytijl : 0,1 variable for shifting i from j to l in period t
Atijl : Cost of shifting i from j to l in period t (where Atijl = 0)
Xtij : 0,1 variable for locating i at j in period t
Ctijkl : Cost of material flow between i located at j and k located
at l in period t
P : Number of periods in the planning horizon
N : Number of departments in the layout

The objective is to minimize the sum of the layout rearrangement cost (first term) and the

material handling costs (second term) over the planning horizon. Constraint (2) requires every

department to be assigned. Constraint (3) requires every location to have a department assigned to it.

Constraint (4) assigns Ytijl to a value of 1 only if a department has been shifted in the period. Both, in

the objective function and in the constraints, decision variables are being multipled with other decision

variables. Thus this formation is nonlinear and can be solved to optimality only for small problems

4. The GA Solutions

11
 In this research, we propose a solution technique based on genetic algorithms (GAs). The
concept of GAs is based on the evolution process that occurs in natural biology. An initial
population of possible solutions (individuals) is generated. Some individuals are selected to be
parents to produce offspring via a crossover operator. All the individuals are then evaluated and
selected based on the concept of survival of the fittest. The process of reproduction, evaluation, and
selection is repeated until a termination criterion is reached. In addition, a mutation operator with
certain probability is appled to the individuals to change their genetic makeup. A mutation process
increases the diversity of the population and ensures an extensive search.
We propose two versions of GAs. They are different in their generational replacement
approaches. The first GA (Method 1) uses a steady state replacement approach, and replace the most
“unfit” individual in each generation. The second GA (Method 2) occasionally will replace a large
number of “unlucky” individuals in a generation.
Similar to Conway and Venkataramanan (1994), we let p[i] denote individual i. When N = 6,
and P = 5, individual i with p[i] = 243156 342651 342615 342615 342516 is interpreted as department
2 is assigned to layout position 1 in the first period and so on.
We will randomly generate a set of initial layouts as the initial population. The selection
process finds two individuals from the population according to the distribution of the relative
strength of the individuals to the entire strength of the population. A single point crossover operator
is likely to produce illegal offspring. We use a point-to-point crossover operator. In a point-to-point
crossover, we no longer cut layout strings randomly. Instead, we cut them at every position, starting
from the first position to the last position of the layout.
To illustrate the idea, consider a six departments, two periods problem. We select 2 layouts
from the population: layout A and layout B. The configurations of layout A and B are shown as
follows:
layout A is p[A] = 123456 234561 and
layout B is p[B] = 214563 345621.
The split will occur at every position, starting from the first position to the last position. If both
layouts (layout A and B) are being split at the first position, sub-strings to the right hand side of the
split of the two layouts are being inter-swapped with each other. That is, two newly layouts will thus
be formed, say layout 1 and layout 2 shown as follows:

p[1] = 114563 345621

p[2] = 223456 234561 (split at the 1st position)

12
 Other layouts produced by splitting and swapping at other positions include:
p[3] = 124563 345621
p[4] = 213456 234561 (split at the 2nd position)

p[5] = 123563 345621

p[6] = 214456 234561 (split at the 3rd position)

p[7] = 123463 345621

p[8] = 214556 234561 (split at the 4th position)

p[9] = 123453 345621

p[10] = 214566 234561 (split at the 5th position)

p[11] = 123456 345621

p[12] = 214563 234561 (split at the 6th position)

p[13] = 123456 245621

p[14] = 214563 334561 (split at the 7th position)

p[15] = 123456 235621

p[16] = 214563 344561 (split at the 8th position)

p[17] = 123456 234621

p[18] = 214563 345561 (split at the 9th position)

p[19] = 123456 234521

p[20] = 214563 345661 (split at the 10th position)

p[21] = 123456 234561

p[22] = 214563 345621 (split at the 11th position).

The point-to-point crossover operator produces 2(NP-1) layouts. As in this case 2(12-1) = 22.
Although many child layouts are illegal, at least one child layout is legal under the point-to-point
crossover. Out of 22 layouts, six child layouts (i.e., p[3], p[4], p[11], p[12], p[21], p[22]) are legal.

13
 It is easy to identify illegal child layouts. An illegal child layout is one with a duplication of
the same department in the same period. A feasibility test is applied each time to eliminate illegal
child layouts.
With a very small probability, mutation may occur in a child layout. Mutation first randomly
chooses a period to undergo mutation. Then within the period chosen, two departments will be
randomly chosen for interchange. For example, mutation applies to p[22] = 214563 345621 and the
resulting string p[22’] = 214563 365421 when the second period and departments 4 and 6 are
chosen.
As the reproduction, mutation, and selection process continues, the population will evolve
over successive generations. The fitness of the best and the average individual in each generation
moves towards the global optimum. The whole terminates when the percentage difference between
the fitness of the best and the average is less than a very small threshold.

14
The flowchart of our first GA implementation (Method 1) is given in Figure 4.

START

Initialize Population

Select two parent layouts

Apply crossover operator

Apply feasibility test

Apply mutation

Replace the worst layout

Satisfy
the stopping
condition?
No
Yes

Find the minimum layout

END

Figure 4: Flowchart of Method 1

15
The second GA implementation (Method 2) is different from Method 1 mainly in the generational
replacement approach. In Method 2, we occasionally replace a large number of “unlucky”
individuals in the population. The flowchart of Method 2 is given in Figure 5. The process in the
inner loop process is similar to the evoluation process in Method 1. The outer loop process allows
the occasional replacement of a number of “unlucky” individuals in the population.

START
Select two parent layouts

Initialize Population
Apply crossover operator

Apply
the evolution process
of Method 1
Apply feasibility test

Outer loop Inner

process Replace loop
some poor layouts process
in the population
Apply mutation

Satisfy
the stopping
condition?
No Replace the worst layout
Yes

Find the minimum layout Satisfy

the stopping
condition?
No
Yes
END

Figure 5: Flowchart of Method 2

5. Experimental Design

The experimental design consists of the following factors and levels:

16
 · Planning horizons: Two different planning horizons of 5 periods and 10 periods

are used. This will give us an idea of the effectiveness of the algorithms over

longer term planning horizons.

· Sizes of layouts (departments or machines): Three layout sizes of 6, 15, and 30

departments are used. This will help us determine the effectiveness of the

algorithms for larger layouts.

Each cell in the experimental design will consist of eight replications. This number has been

shown to be sufficient for low anderrors for similar problems (Balakrishnan et al.). For each

problem, the sum of the flows in a layout during a period was made constant during the entire

horizon. This is to prevent any period from biasing the others. The average shifting cost for a

department was set to be 15% of the average material handling cost for the department. The five

period problems use the first five periods of data from the 10 department problems. In each problem,

some departments had high inflows as compared to the others. These high inflow departments were

changed in every period to simulate demand changes. An example of this can be seen in Figure 3.

6. Computational Results

We have done a simple experiment to test our implementations with different parameters. For

example, we ran experimental runs with population sizes of 100, 500, and 1000. We also performed

experimental runs with mutation probabilities of 0.001, 0.002, and 0.003. In addition, we also ran

experimental runs with the threshold values for the stopping conditions at 1%, 5%, and 10%. Based on

the results, we choose the following parameter values: population size = 500, mutation probability =

0.002, and the threshold value for the stopping condition is 5%.

The results are shown in Tables 1 - 6. There are two type of tables; the first type of table gives

the 5 period, and the second type gives the 10 period results. Each table gives the minimal cost solution

for each of the 8 problems (data 1 to data 8) using the following methods:

· Genetic Search by Conway and Venkataramanan (1994): GEN S

17
 · GA : Method 1

· GA : Method 2

The computational results are summarised in Tables 1 to 6. In general, Method 2 consistently

outperforms Method 1 in our study. The difference between Method 1 and Method 2 is mainly in the

generational replacement approaches. The computational results seem to indicate that although the

point-to-point crossover operator may avoid illegal child layouts, it may not guarantee an extensive

search in the search space. The generational replacement approach in Method 2 helps increase

population diversity and address the limitation of the crossover operator.

In Tables 1 and 2 (i.e., 6 departments, 5 and 10 periods respectively), Method 2 performs worse

than GEN S in one problem case. In Tables 3 and 4 (i.e., 15 departments, 5 and 10 periods

respectively), GEN S performs better than Method 2 in 6 problem cases. Finally, in Tables 5 and 6 (i.e.,

30 departments, 5 and 10 periods respectively), Methods 2 performs better in all cases.

7. Conclusion

In this paper, we study the applicability of genetic algorithms (GAs) in solving the dynamic

layout problem. Two versions of the GA implementation have been proposed. Our GA approaches

differ from the existing implementation in two ways:

· We employ a new crossover operator to avoid illegal offspring.

· We adopt a different generational replacement approach to increase population diversity.

With the new feature, one of our implementations is able to generate solutions better than an existing

implementation in our computation study.

References

Armour G.C., and Buffa E.S., “A heuristic algorithm and simulation approach to relative allocation of
facilities”, Management Science, 9, 2, 294-300, 1963.
Balakrishnan J., “The dynamics of plant layout”, Management Science, 39, 5, 654-655, 1993.
Balakrishnan J., Jacobs R.F., Venkataramanan M.A., “Solutions for the constrained dynamic facility
layout problem”, European Journal of Operations Research, 57, 280-286, 1992.
Ballou R.H., “Dynamic warehouse location analysis”, Journal of Marketing Research, 5, 271-276,
1968.

18
Bellman R.E., and Dreyfus S.E., Applied Dynamic Programming, Princeton, New Jersey: Princeton
University Press 1962.
Conway D.G., and Venkataramanan M.A., “Genetic search and the dynamic facility layout problem”,
Computers & Operations Research, 21, 8, 955-960, 1994.
Koopmans T.C., and Beckman M., “Assignment problems and location of economic activities”,
Econometrika, 25, 53-76, 1957.
Kusiak A., and Heragu S.S., “The facility layout problem”, European Journal of Operational Research,
29, 229-251, 1987.
Lodish L.M., “Computational limitations of dynamic programming for warehouse location”, Journal of
Marketing Research, 7, 262-263, 1970.
Mote J., Murthy I., and Olson D.L., “On solving singly constrained shortest path problems by
generating pareto-optimal paths parametrically”, Working Paper, Louisiana State University, Baton
Rouge, Louisiana, 1988.
Page A.L., “PDMA new product development survey: performance, and best practices”, Paper
presented at PDMA Conference, Chicago, Nov. 13, 1991.
Rosenblatt M.J., “The dynamics of plant layout”, Management Science, 32, 1, 1986, 76-86.
Sahni S., and Gonzales T., “P-complete approximation problem” Journal of the ACM, 23, 3, 555-565,
1976.
Sweeney D.S., and Tatham R.L., “An improved long run model for multiple warehouse location”,
Management Science, 22, 7, 758-758, 1976.
Tompkins J.A., and White J.A., Facilities Planning, Wiley:New York, 1984
Urban T.L., “A heuristic for the dynamic facility layout problem”, IIE Transactions, 25, 4, 57-63,
1993.

19
Table 1
6 Departments and 5 periods

1 2 3 4 5 6 7 8
GEN S 108976 105170 104520 106719 105628 105606 106439 104485
Method 1 109935 106128 105385 107407 106523 104702 108478 107387
Method 2 106419 104834 104320 106515 105628 104053 106978 103771

Table 2
6 Departments and 10 periods

1 2 3 4 5 6 7 8
GEN S 218407 215623 211028 217493 215363 215564 220529 216291
Method 1 220236 215151 215307 219872 217615 217808 223727 223729
Method 2 214397 212138 208453 212953 211575 210801 215685 214657

Table 3
15 Departments and 5 periods

1 2 3 4 5 6 7 8
GEN S 504759 514718 516063 508532 515599 509384 512508 514839
Method 1 528932 531854 544481 539147 540702 532387 532622 538183
Method 2 511854 507694 518461 514242 512834 513763 512722 521116

Table 4
15 Departments and 10 periods

1 2 3 4 5 6 7 8
GEN S2 1055536 1061940 1073603 1060034 1064692 1066370 1066617 1068216
Method 1 1079510 1076619 1091916 1083110 1089044 1078395 1086191 1084743
Method 2 1047596 1037580 1056185 1026789 1033591 1028606 1043823 1048853

Table 5
30 Departments and 5 periods

1 2 3 4 5 6 7 8
GEN S2 632737 647585 642295 634626 639693 637620 640482 635776
Method 1 669131 669669 671745 664838 660862 670429 668859 674047
Method 2 611794 611873 611664 611766 604564 606010 607134 620183

Table 6
30 Departments and 10 periods

20
 1 2 3 4 5 6 7 8
GEN S 1362513 1379640 1365024 1367130 1356860 1372513 1382799 1383610
Method 1 1365904 1369680 1357917 1361263 1354237 1358267 1375524 1375393
Method 2 1228411 1231978 1231829 1227413 1215256 1221356 1212273 1245423

21